math-stats-ml 0.0.9

Python package for Mathematical Statistics with a View Toward Machine Learning, by John Myers

Utility functions for "Mathematical Statistics with a View Toward Machine Learning", by John Myers

A Python package for all utility functions used in the book and programming assignments. To install, do the usual:

pip install math_stats_ml

All other materials are contained here.

Table of contents:

1. gd submodule: Gradient descent utilities
   • GD_output class: Container class for output of algorithms
   • GD function: Gradient descent
   • SGD function: Stochastic gradient descent
   • plot_gd function: plot the output of gradient descent
   • plot_sgd function: plot the output of stochastic gradient descent

gd submodule: Gradient descent utilities

Contains all utilities for the gradient descent algorithms used in the book.

GD_output class: Container class for output of algorithms

class GD_output

A class holding the outputs of both the gradient descent (GD) and stochastic gradient descent (SGD) algorithms. All attributes below are optional and default to None.

Attributes

Name	Type	Description
parameters	dict	A dictionary containing the parameters of the objective function passed to either GD or SGD. Each value in the dictionary is a tensor whose zero-th dimension indexes the number of gradient steps.
per_step_objectives	torch.Tensor	A tensor containing the running objective values, per gradient step.
per_epoch_objectives	torch.Tensor	A tensor containing the running mean objective values, per epoch.
epoch_step_nums	torch.Tensor	A tensor containing the number of each gradient step on which an epoch begins/ends.
grad_steps	iter	An iterable ranging from $0$ to one less than the total number of gradient steps. (This is convenient for plotting purposes.)
lr	float	Learning rate.
num_steps	int	Number of gradient steps to run the gradient descent (GD) algorithm.
decay_rate	float	Learning rate decay.
batch_size	int	Mini-batch size for the stochastic gradient descent (SGD) algorithm.
num_epochs	int	Number of epochs for the stochastic gradient descent (SGD) algorithm.
max_steps	int	Maximum number of gradient steps after which we terminate the stochastic gradient descent (SGD) algorithm.
type_flag	str	Either None or gd. In the latter case, indicates whether the GD_output object was obtained from the gradient descent (GD) algorithm.

GD function: Gradient descent

GD(J, init_parameters, lr, num_steps, decay_rate=0)

Implementation of gradient descent. The notation below is intended to match the notation in the description in the book.

Output

The output type is an object of the GD_output class.

Parameters

Name	Type	Description
J	function	Objective function to be minimized. The parameters of the function are either a single tensor or a dictionary of tensors (in the case that the parameters fall into natural groups, e.g., weights and biases).
init_parameters	torch.Tensor or dict	Initial parameters.
lr	float	Learning rate, corresponding to $\alpha$ in the book.
num_steps	int	The number of gradient steps after which the algorithm should halt, corresponding to $N$ in the book.
decay_rate	float	Learning rate decay, corresponding to $\beta$ in the book. Defaults to 0.

SGD function: Stochastic gradient descent

SGD(g, init_parameters, X, lr, batch_size, num_epochs, y=None, decay_rate=0, max_steps=-1, shuffle=True, random_state=None)

Implementation of stochastic gradient descent. The notation and terminology below is intended to match the book.

Output

The output type is an object of the GD_output class.

Parameters

Name	Type	Description
g	function	Target function for the algorithm. (The notation and terminology is intended to match the book---see the link above.) The call signature of the function is of the form g(parameters, x) or g(parameters, x, y), where x is a feature vector and y is an (optional) ground truth label of a single instance in the dataset, and parameters is either a single parameter tensor or a dictionary of parameter tensors (in the case that the parameters fall into natural groups, e.g., weights and biases). We assume that g is "vectorized," so that it may accept a design matrix X in place of x and an entire vector of ground truth labels for y.
init_parameters	torch.Tensor or dict	Initial parameters.
X	torch.Tensor	Design matrix. The rows are the feature vectors that are fed into the function g.
lr	float	Learning rate, corresponding to $\alpha$ in the book.
batch_size	int	Mini-batch size, corresponding to $k$ in the book.
num_epochs	int	The number of epochs after which the algorithm should halt, corresponding to $N$ in the book.
y	torch.Tensor	Vector of ground truth labels for the data in the design matrix X. Optional, defaults to None.
decay_rate	float	Learning rate decay, corresponding to $\beta$ in the book. Defaults to 0.
max_steps	int	Maximum number of gradient steps after which the algorithm should halt. Defaults to -1, in which case the algorithm will complete all num_epochs many epochs.
shuffle	bool	Determines whether to shuffle the dataset before looping through an epoch. Defaults to True.
random_state	int	If not None and shuffle=True, random seed to be passed to torch.manual_seed. Defaults to None.

plot_gd function: plot the output of gradient descent

plot_gd(gd_output, w=5, h=4, plot_title=True, parameter_title=True, show_xlabel=True, xlabel='gradient steps', show_ylabel=True, ylabel='surprisal', alpha=1, color=None, ax=None)

Later...

plot_sgd function: plot the output of stochastic gradient descent

plot_sgd(sgd_output, w=5, h=4, plot_title=True, parameter_title=True, show_step=True, show_epoch=True, show_xlabel=True, xlabel='gradient steps', show_ylabel=True, ylabel='cross entropy', legend=True, per_step_alpha=0.25, per_step_color=None, per_step_label='cross entropy per step', per_epoch_color=None, per_epoch_label='mean cross entropy per epoch', s=10, ax=None)

Later...